Research On Cross-domain Target Detection Method Of Underwater Vehicle In Complex Environment

Seafood has high medicinal value and nutritional value,but the traditional way of artificial fishing has seriously restricted the market development of seafood.With the increasing maturity of underwater robot technology,the use of underwater robots for intelligent fishing has become a trend.Underwater vehicle often faces the problem of changing or replacing the application scenarios when performing the task of capturing marine organisms,which will lead to the problem of domain offset between the actual application scenarios of target detection task and the training data due to the different data distribution.Common target detection models have the problem of poor detection effect due to domain offset,which seriously affects the ability of autonomous operation of underwater vehicles.The domain adaptive method can apply the knowledge learned in the scene with labeled information to the new scene without labeled information through transfer learning,which effectively alleviates the problem of accuracy decline caused by domain offset and enhances the operation ability of underwater vehicles in the new scene.In this paper,the application of domain adaptive method in underwater scene is studied.The main research contents are as follows:In order to solve the problem that it is difficult to align the feature representation of sparse foreground objects and the feature representation of multi-scale objects when the image-level domain adaptation module is oriented to underwater scenes,a multi-scale image-level domain adaptation module based on dual attention is proposed.The module uses dual attention mechanism to aggregate the features of sparse foreground regions,and uses multi-scale domain adaptive module to retain more small-scale target feature information,which effectively enhances the degree of image-level feature alignment.The experiment of cross-domain target detection is set up by combining three cross-domain data sets,namely,pool scene to real ocean scene,different sea scene and synthetic data set to real ocean scene.The experimental results show that the proposed method achieves better detection results on three data sets,and effectively alleviates the problem of domain offset of image level in underwater capture scene transformation,The degree of image-level feature alignment is enhanced.When the instance-level domain adaptation module is applied to underwater scenes,the instance-level features of the target are difficult to detect and the instance-levels features of the background are invalid,which affect the alignment accuracy of instance-level feature representation.To solve this problem,a classified-regularization-based instance domain adaptation module is proposed.The module distinguish that instances which are difficult to detect,the foreground object instance and the invalid background instances according to the difference between the image-level class prediction and the instance-level clas prediction,and then different weights are applied to different instance-level features through a weighting factor,and the instance-level feature representations with different weights are added to domain adaptive training,so that the alignment fineness of the instancelevel feature representations is enhanced.Experimental results on three cross-domain data sets show that the proposed method achieves better detection results on three cross-domain data sets,and effectively enhances the fineness of instance-level feature alignment in underwater grasping scenes.